We are frequently asked whether companies should build their own AI LLM software in house or buy something from a vendor. Building software in house is expensive and can come with a lot of headaches. But it can also create a lot of cost-savings for some organizations. While there are other factors an organization may consider when deciding to build or buy, those can often be mitigated. For instance, OpenAI's ChatGPT for Enterprise is SOC II compliant and has at-rest and in-transit encryption.
With that said, we conducted both an ad-hoc cost analysis and a quantitative algebraic comparison analysis to help companies decide whether to build or buy their AI LLM software.
In House Costs
According to Glassdoor at the time of this writing, the average data scientist salary is $152,279 USD. That’s equivalent to 1,001 annual subscriptions to Silatus Pro+ and before you factor in variable server costs and GPU and/or OpenAI API costs. Add a UI/UX designer ($92,258) and a small team of four software engineers ($139,721) for maintenance and new features, and your total cost balloons to at least $384,258 per year.
To make fixed costs a non-factor and clearly justify the cost of building in-house, fixed costs should probably be about an order of magnitude (10x) lower than variable costs. If you don't like that approach, keep reading. We'll take a more precise approach in a moment.
Let's start the basic calculation. At scale, the average LLM user consumes between $70 - $150 in compute resources per year. For big data processing, this number can get much higher, but we’ll use the lowest number as a conservative estimate.
In order for your variable costs to equal your fixed costs in this scenario, you need at least 5,489 employees actively using your LLM software. Increase that by an order of magnitude, and you need 54,890 employees actively using your LLM software for your fixed costs to be a non-factor in your build vs. buy cost analysis.
Build vs Buy Cost Estimation
This is great, but it doesn't precisely tell us whether a company would derive more cost-savings from building in-house versus leveraging a third-party vendor. We can also do an algebraic analysis to try to extract a more precise value.
An annual Silatus Pro+ subscription is $144 per year. Variable costs are assumed to be $70 per user per year and fixed costs are assumed to be $384258 divided by the number of employees actively using the software. To get total in-house cost, we can use the formula: TotalCost = VariableCost + FixedCost / Users
If x = active employee users and y = total cost per year, we can say: y = 70 + 384258 / x
This equation is graphed below:
To find the “midpoint” of this function, where total cost per user is equal to the number of users, we find where the line y = x intersects with our equation. We can do this by simply substituting y for x, getting us x = 70 + 384258 / x. This results in x = 655 users and since y = x, $655 per user per year. That’s expensive!
Next, let’s find the point where a Silatus Pro+ annual subscription is equal to the total cost. Let’s say the Pro+ subscription is z = $114 per user per year. We want to find the value for x where y = z. We get the equation 114 = 70 + 384258 / x. If we graph this equation, we get a vertical line intersecting our original graph at x = 8,733 users.
So, conservatively, assuming you could achieve similar or better results to ours, you would need 8,733 active employee users just for your costs to be equivalent to purchasing a tool like Silatus. To extract a 2x cost-savings, you need over 17,000 active employee users. If you have that many employees that you expect to use your software, then by all means, consider building in-house.
Yet still, consider this. As third-party products scale and mature to more and more engineers and better techniques, fixed costs relative to building in house will increase. That very conservative 8,733 user number will only go up because vendors like us plan to deliver far more than $384,258 in human engineering value over time.
As companies are faced with the choice of developing AI LLM software internally or procuring it from vendors, it's crucial for them to weigh the initial expenditures against long-term value and extensibility. Our analysis demonstrates that while creating software in-house provides a certain level of control and potential cost benefits, these gains are significant only when there's a large-scale employee base of users.
Moreover, as third-party suppliers like us persistently scale up and enhance their solutions, the relative fixed expenses of in-house development are likely to increase, making the option of external solutions increasingly cost-effective over time. Thus, unless a company has a massive user base and is ready to tackle the ongoing costs and hurdles related to internal development, capitalizing on external vendors emerges as a more budget-friendly and efficient strategy for most businesses.
While most of this post was written by a human (LLMs are bad at math), the conclusion was written by Silatus AI.